[Bloat] The NetEm qdisc does not work in conjunction with other qdiscs

[Dave Taht]

To add to discrepancies tbf so far as I remember broke up gso
superpackets into packets, htb did not.

How all the aqms behave with superpackets is insufficiently explored. This
is why cake made breaking
up gso packets into regular packets the default.


On Fri, Oct 25, 2024 at 3:47 PM Joerg Deutschmann via Bloat <
bloat at lists.bufferbloat.net> wrote:

> Thanks Max and Stephen,
>
> >>  From my experience (experimented with it last year), it still behaves
> >> weirdly. You can use htb for the shaping and you can create delay using
> >> netem by using it on another (virtual) host on a link that does not have
> >> any other qdiscs and where the link is not the bottleneck.
> Yes, I did some quick tests and saw that this works:
> > sudo tc qdisc add dev eth0 root handle 2: tbf rate 200Mbit burst 4542
> limit 500000
> > sudo tc qdisc add dev eth0 parent 2: handle 3: fq_codel
> but this does not work:
> > sudo tc qdisc add dev eth0 root handle 1: netem delay 20ms limit 10000000
> > sudo tc qdisc add dev eth0 parent 1: handle 2: tbf rate 200Mbit burst
> 4542 limit 500000
> > sudo tc qdisc add dev eth0 parent 2: handle 3: fq_codel
>
> Adding netem delay to a separate host or adding it to the ingress qdisc
> (as described by Dave) does the trick.
>
>
> > Yes, netem has expectations about how inner qdisc behaves,
> > and other qdisc used as inner have expectations about how/when packets
> > are sent. There is a mismatch, not sure if it is fixable with in the
> > architecture of how Linux queue disciplines operate.
> For the sake of completeness, here's the quote from the netem man page:
> > Combining netem with other qdisc is possible but may not always work
> because netem use skb control block to set delays.
> :-)
>
> >
> > The best way is to use netem on an intermediate hop and not expect
> > it to work perfectly when used on an endpoint. The same is true of
> Dummynet
> > and other network emulators; they are uses as man-in-the-middle systems.
> Still I guess it can be tricky if netem delay + tbf shaping + fq_codel
> shall be used altogether.
>
> This paper mentions that the order tbf/netem or netem/tbf matters
> (section 4.1.1), but does not mention fq_codel:
> https://atlas.cs.uni-tuebingen.de/~menth/papers/Menth17-Sub-2.pdf
> _______________________________________________
> Bloat mailing list
> Bloat at lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/bloat
>


-- 
Dave Täht CSO, LibreQos
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.bufferbloat.net/pipermail/bloat/attachments/20241025/9db99043/attachment-0001.html>